Skip to content

RAGFlow is an agent-native database: a Retrieval-Augmented Generation (RAG) engine that grew into context infrastructure.

6.3 KB
Raw Blame
//
//  Copyright 2026 The InfiniFlow Authors. All Rights Reserved.
//
//  Licensed under the Apache License, Version 2.0 (the "License");
//  you may not use this file except in compliance with the License.
//  You may obtain a copy of the License at
//
//      http://www.apache.org/licenses/LICENSE-2.0
//
//  Unless required by applicable law or agreed to in writing, software
//  distributed under the License is distributed on an "AS IS" BASIS,
//  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
//  See the License for the specific language governing permissions and
//  limitations under the License.
//

package indexdoc

// positionsToIntMatrix groups a flat position array by five and truncates each
// value to int — the stored shape of position_int. It returns nil for an empty
// array or one whose length is not a multiple of five: malformed input is
// dropped, never reinterpreted.
func positionsToIntMatrix(positions []float64) [][]int {
	if len(positions) == 0 || len(positions)%5 != 0 {
		return nil
	}
	matrix := make([][]int, 0, len(positions)/5)
	for i := 0; i < len(positions); i += 5 {
		matrix = append(matrix, []int{
			int(positions[i]), int(positions[i+1]), int(positions[i+2]),
			int(positions[i+3]), int(positions[i+4]),
		})
	}
	return matrix
}

// positionRowsFromAny normalizes the generic []any shape produced by
// ChunkDoc.ToMap (JSON decoding into `any`) into [][]any. It accepts a list of
// rows (each []any/[]float64), a single flat 5-tuple, or a flat sequence of
// multiple 5-tuples (TCADP/external producers).
//
// Element 1 distinguishes the shapes: a row list has another row (or an
// invalid entry) there, while a flat sequence has a numeric coordinate. A flat
// sequence is then grouped by five; a length that is not a multiple of five is
// malformed, matching positionsToIntMatrix.
func positionRowsFromAny(v []any) ([][]any, bool) {
	if len(v) == 0 {
		return nil, false
	}
	if len(v) == 1 {
		if row, ok := positionRowFromAny(v[0]); ok {
			return [][]any{row}, true
		}
	}
	if len(v) >= 2 {
		_, secondIsRow := positionRowFromAny(v[1])
		if secondIsRow || v[1] == nil {
			_, firstIsRow := positionRowFromAny(v[0])
			if firstIsRow || v[0] == nil {
				rows := make([][]any, 0, len(v))
				for _, e := range v {
					if row, ok := positionRowFromAny(e); ok {
						rows = append(rows, row)
					}
				}
				return rows, len(rows) > 0
			}
		}
	}
	if len(v)%5 != 0 {
		return nil, false
	}
	rows := make([][]any, 0, len(v)/5)
	for i := 0; i < len(v); i += 5 {
		rows = append(rows, v[i:i+5])
	}
	return rows, len(rows) > 0
}

func positionRowFromAny(e any) ([]any, bool) {
	switch row := e.(type) {
	case []any:
		return row, true
	case []float64:
		out := make([]any, len(row))
		for i, value := range row {
			out[i] = value
		}
		return out, true
	default:
		return nil, false
	}
}

// flattenMixedPositions expands a [][]any position matrix into the flat
// [page,left,right,top,bottom,...] form addPDFPositions expects. This is the
// shape left by the title-family merge when a page cell is a list (or rows are
// otherwise non-float); each page in a list becomes its own scalar tuple so
// page_num_int/top_int can be built without discarding the whole matrix.
func flattenMixedPositions(rows [][]any) ([]float64, bool) {
	flat := make([]float64, 0, len(rows)*5)
	for _, row := range rows {
		if len(row) < 5 {
			continue
		}
		left, lok := positionNumber(row[1])
		right, rok := positionNumber(row[2])
		top, tok := positionNumber(row[3])
		bottom, bok := positionNumber(row[4])
		if !lok || !rok || !tok || !bok {
			continue
		}
		pages, ok := positionPages(row[0])
		if !ok {
			continue
		}
		for _, page := range pages {
			flat = append(flat, page, left, right, top, bottom)
		}
	}
	return flat, len(flat) > 0
}

func positionPages(raw any) ([]float64, bool) {
	switch v := raw.(type) {
	case []any:
		pages := make([]float64, 0, len(v))
		for _, e := range v {
			page, ok := positionNumber(e)
			if !ok {
				return nil, false
			}
			pages = append(pages, page)
		}
		return pages, len(pages) > 0
	default:
		page, ok := positionNumber(v)
		if !ok {
			return nil, false
		}
		return []float64{page}, true
	}
}

func positionNumber(raw any) (float64, bool) {
	switch v := raw.(type) {
	case float64:
		return v, true
	case float32:
		return float64(v), true
	case int:
		return float64(v), true
	case int32:
		return float64(v), true
	case int64:
		return float64(v), true
	default:
		return 0, false
	}
}

// addPDFPositions stores PDF layout boxes: positions is a flat []float64
// grouped by five as [page, left, right, top, bottom] and is written to
// page_num_int / top_int / position_int. Page numbers are ALREADY 1-indexed on
// entry — the 0→1 conversion happens once, at the parser boundary
// (normalizePDFPageNumber for the DeepDoc PDF path; TCADP writes 1-indexed
// directly) — so this is a passthrough and must NOT add +1, or the PDF path
// would double-increment.
//
// Mirrors Python: rag.nlp.add_positions() (Python adds +1 because its callers
// feed 0-indexed values; the Go pipeline normalizes earlier).
func addPDFPositions(chunk map[string]any, positions []float64) {
	matrix := positionsToIntMatrix(positions)
	if matrix == nil {
		return
	}
	pageNumInt := make([]int, 0, len(matrix))
	topInt := make([]int, 0, len(matrix))
	for _, box := range matrix {
		pageNumInt = append(pageNumInt, box[0])
		topInt = append(topInt, box[3])
	}
	chunk["page_num_int"] = pageNumInt
	chunk["top_int"] = topInt
	chunk["position_int"] = matrix
}

// addSpreadsheetPositions stores spreadsheet cell ranges: positions is a flat
// []float64 grouped by five as [sheet, rowStart, rowEnd, colStart, colEnd] and
// is written to position_int only — that field is the document preview's
// coordinate carrier (the chunk API serves "positions" from it, and the Excel
// preview reads the five values positionally as sheet / rows / columns).
//
// page_num_int and top_int are PDF layout fields: deriving them from these
// tuples would store a sheet number as a page and a column as a top, and would
// clobber the row index the QA chunker wrote to top_int (the same field
// Python's beAdoc/beAdocDocx set from row_num for its spreadsheet chunks,
// which never carry positions at all).
func addSpreadsheetPositions(chunk map[string]any, positions []float64) {
	if matrix := positionsToIntMatrix(positions); matrix != nil {
		chunk["position_int"] = matrix
	}
}